Background: Early-onset emotional and behavioral difficulties (EBD) in children are a significant concern. Yet early identification remains a major challenge, especially for children with emerging or sub-clinical issues who may be overlooked and are at high risk for worsening difficulties if their vulnerabilities go unrecognized.
Objective: This study aims to develop and validate an interpretable machine learning model to predict different risk levels of EBD in young children. By integrating observable and objective interaction factors, the model seeks to facilitate the early detection and monitoring of children's mental health.
Methods: A cross-sectional study was conducted among 1,034 mothers with children aged 1-4 years, recruited from 10 community vaccination centers in Zhejiang, China. Socio-demographic, interparental and parent-child interaction factors, and child EBD risks were assessed. Six machine learning algorithms were applied to develop predictive models, which were interpreted using SHapley Additive exPlanations.
Results: The Random Forest model achieved the best overall performance, demonstrating strong discriminative power (AUC = 0.892) with well-balanced sensitivity (0.72) and specificity (0.74). The model was highly effective at identifying both high-risk (AUC = 0.90) and low-risk (AUC = 0.91) children. "Parents being easily provoked by children's negative behaviors" was the most important feature in all models. Other important features included difficulty forming warm parent-child bonds, interparental problems due to emotional volatility, fewer child smiles during interactions, and intimate partner violence.
Conclusions: This study suggests the optimal models hold promise for screening young children at different EBD risk levels, not only those with clear difficulties or well-adjusted profiles, but also those with latent vulnerabilities who may benefit from early support. Preventive strategies should emphasize family interaction patterns to improve early identification and subsequent intervention.
